FLUX 3 Image launch discount: a 40-image test plan before it ends

If you want to judge FLUX 3 Image on your own images, the cheapest time is now: a 40-image test at 1K costs about $0.96 at the discounted rate BFL's pricing calculator shows ($0.024 per image) and about $1.92 at the full rate ($0.048). Press coverage reports the launch discount as 50 percent through October 8, 2026; BFL's own docs pages I read do not state the end date, so confirm it in your BFL dashboard before you budget around it.

FLUX 3 Image is not a model Sume lists. The Sume image docs name FLUX.2 Pro and FLUX.2 Flex under the Black Forest Labs prefix, plus GPT Image 2.5, Ideogram 4.5, Nano Banana and others, and say nothing about FLUX 3. So the practical split is: test FLUX 3 on BFL's API while it is cheap, and run the same prompts through what you can call on Sume today as your baseline.

What the discount is worth in dollars

The saving is real but small in absolute terms. At 1K the difference is $0.024 per image, so even 1,000 test images only save $24. The discount matters more if you plan a large evaluation at 2K or 4K, where BFL's docs list $0.100 and $0.607 per image at full price. I could not find the discounted 2K and 4K figures on a BFL page, so I do not quote them.

Design the test so it can change your mind

Forty images is enough for a first pass if each one tests a feature you would pay for. Pick the things FLUX 3 Image claims that your current model cannot do: multi-reference edits with several inputs, a layout described with bounding boxes, an edit that must leave the rest of the picture alone, and a tall or wide aspect ratio.

Keep every variable but the model fixed. Same prompt text, same reference files, same aspect ratio, and run each prompt twice so a lucky seed does not decide the verdict. Have two people mark each result pass or fail on a short written checklist before they see which model made it.

• 10 prompts that need two or more references (a product plus a scene, a person plus an outfit)
• 10 edits where only one region may change; mark any pixel drift as a fail
• 10 layout prompts with named positions (left third, top banner)
• 10 plain text-to-image prompts at the ratio you actually publish

Run the baseline on Sume

Sume's Image API takes a catalog model id and returns hosted URLs, with the billed amount in usage.cost. Per the docs, POST /v1/images blocks for up to 30 seconds and answers 200 with the image, and anything slower returns 202 with a job to poll, so the script below treats 202 as a note to read the job later. Check each model's price and parameters with GET /v1/images/models before a bigger run.

The same three prompts go to FLUX.2 Pro and GPT Image 2.5, which gives you a spend figure and two sets of outputs to line up against the FLUX 3 results.

import os
import requests

PROMPTS = [
    "Product still: a matte black water bottle on wet slate, hard side light, 4:5",
    "Poster: the words OPEN LATE in cut paper letters, one line, centered",
    "Interior: a small cafe corner, morning light, no people",
]
MODELS = ["black-forest-labs/flux.2-pro", "openai/gpt-image-2.5"]

def generate(payload):
    r = requests.post(
        "https://api.sume.com/v1/images",
        json=payload,
        timeout=60,
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    )
    if r.status_code == 202:
        return {"data": [{"url": "(202 job, read " + r.json()["data"]["status_url"] + ")"}], "usage": {"cost": 0}}
    r.raise_for_status()
    return r.json()

total = 0.0
for model in MODELS:
    for prompt in PROMPTS:
        out = generate({"model": model, "prompt": prompt, "aspect_ratio": "1:1"})
        total += out["usage"]["cost"]
        print(model, out["data"][0]["url"])
print(f"baseline spend: ${total:.4f}")

When not to chase the discount

Skip it if your pipeline already works on a model Sume serves and your blocker is not quality. BFL has said an open-weight FLUX 3 version is expected, without a date, so a self-hosting plan should not hinge on a one-week price window. If the test shows a clear win on your edit-heavy prompts, you will have the evidence to ask for FLUX 3 on a catalog; if it is a tie, you saved yourself a migration.

For how FLUX 3's resolution tiers line up against what Sume lists, see FLUX 3 Image price by tier versus Sume pricing lookup. The model list itself is in the Image API docs.